多摄像头非刚体目标检测与空间定位系统

发布时间：2018-09-19 12:40

【摘要】：随着视频监控系统的迅速普及,计算机视觉正逐渐为大众所熟知,特别是计算机视觉领域的运动目标检测与定位技术,近年来受到越来越多的重视,已广泛应用于安防监控、入侵检测、无人驾驶汽车等领域。传统的视频监控系统需要较多的人工参与,无法应对越来越复杂多变的监控环境,而基于视频图像的智能视频监控系统不需要或者只需要很少的人工参与,可以同时对多个场景进行监控,能够在极短的时间内分析、发现监控中的异常行为并实时发出告警。本文对视频监控中的核心目标检测技术进行了深入研究,分析当今主流的目标检测算法,并针对具体的应用场景,使用了多种目标检测算法构建目标检测系统,在多台摄像机上应用该系统,构建了目标空间定位系统。本文的主要内容有:1.提出一种基于帧间差分和相似度校验的多目标检测方法。该方法采用帧间差分法得到差分图像,经过形态学处理得到目标轮廓,针对非刚体目标形变引起的目标分裂等问题,采用多种方法进行合并,通过计算区域相关性辨别多目标,可以快速确定目标在图像中的粗略位置,同时具有较高的运行效率,为目标的精确检测提供位置参考。2.研究基于可变形部件模型的目标检测算法,在差分算法得到目标粗略位置的基础上,通过使用该算法可以获得指定类型目标的精确位置。可变形部件模型是一种基于目标特征统计学习的检测算法,该方法使用HOG描述子作为模型特征,具有良好的几何和光学转换不变性;同时由于形变模型的引入,使得该方法对于物体的非刚体形变具有极好的鲁棒特性,特别使用于行人等非刚体目标检测的应用场景。3.研究基于图像的目标空间定位方法,在两点前向交会的基础上,提出了一种基于联合概率分布的多路交会算法,使用概率模型根据多台摄像机的检测结果得到目标在空间位置的估计值,这种方法比传统的多次使用两路交会然后取几何中心的方法可靠性和精度都有了极大提高;通过数学模型对定位系统在水平方向和垂直方向的最小理论误差进行了分析,得出了系统实施环境下定位的理论精度极限。
[Abstract]:With the rapid popularization of video surveillance system, computer vision is gradually becoming known to the public, especially in the field of computer vision moving object detection and location technology, in recent years, more and more attention has been paid to, has been widely used in security surveillance, intrusion detection, driverless vehicles and other fields. Traditional video surveillance system needs more. Artificial participation can not cope with the increasingly complex and changeable monitoring environment. However, intelligent video surveillance system based on video images does not need or only need a small amount of manual participation. It can simultaneously monitor multiple scenes. It can analyze in a very short time, discover abnormal behavior in monitoring and give real-time alarm. In this paper, the core target detection technology in frequency monitoring is deeply studied, and the main target detection algorithms are analyzed. Aiming at the specific application scenarios, a target detection system is constructed by using a variety of target detection algorithms, and a target location system is constructed by using the system on multiple cameras. The main contents of this paper are as follows: 1. A multi-target detection method based on inter-frame difference and similarity checking is proposed. The difference image is obtained by inter-frame difference method, and the object contour is obtained by morphological processing. In order to solve the problem of target splitting caused by non-rigid object deformation, a variety of methods are used to merge. The multi-target can be identified quickly by calculating the region correlation. The rough position of the target in the image is determined, and the operation efficiency is high, which can provide a reference for the accurate detection of the target. 2. The target detection algorithm based on deformable component model is studied. On the basis of the rough position of the target obtained by the difference algorithm, the precise position of the specified type of target can be obtained by using this algorithm. The shape component model is a detection algorithm based on target feature statistical learning, which uses HOG descriptors as model features and has good geometric and optical transformation invariance. At the same time, due to the introduction of deformation model, this method has excellent robustness for non-rigid deformation of objects, especially for pedestrians and other non-rigid deformation. 3. The application scene of object detection is studied. Based on the two-point forward rendezvous, a multi-path rendezvous algorithm based on joint probability distribution is proposed. The estimation value of object position in space is obtained by using probability model according to the detection result of multiple cameras. This method is more than the traditional method. The reliability and accuracy of the method of using two paths to intersect and then take the geometric center are greatly improved. The minimum theoretical error of the positioning system in the horizontal and vertical directions is analyzed by mathematical model, and the theoretical accuracy limit of the positioning system in the implementation environment is obtained.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TP391.41;TN948.6

【相似文献】